A credential-security ISV
AWS-validated reference — full details available to AWS Partner Validation or on request. (Customer name held on file with VeUP.)
VeUP advised a credential-security ISV through a production agentic build on Amazon Bedrock — Bedrock Agents + Knowledge Bases RAG + Guardrails on Anthropic Claude (3.5 Sonnet → Haiku). Guardrails enforce prompt-injection protection; a benchmarked Haiku/Sonnet/Nova selection powers it.
The challenge
the customer set out to add conversational AI to a security product — the category where the trust bar is highest. It needed AI agents that could answer customer product questions with session continuity and drive an interactive onboarding flow (creating users and credentials directly in the admin console) without becoming a new attack surface. For a credential-security vendor, prompt-injection resistance and a defensible model-selection rationale were not optional, and latency-sensitive steps run in the user's path so the agents had to stay fast and cost-controlled.
The solution
VeUP advised the customer hands-on through a production agentic build on Amazon Bedrock, using Bedrock Agents to orchestrate a multi-step flow — session memory, Bedrock Knowledge Bases for retrieval-augmented generation over a vector store of product Q&A and files, response generation, and output streaming. Amazon Bedrock Guardrails were configured to block prompt-injection and related attacks — the load-bearing control for a security ISV. The agents launched on Anthropic Claude 3.5 Sonnet; VeUP advised moving latency-sensitive steps to Claude Haiku (faster, cheaper) while reserving Sonnet for deep-reasoning steps. the customer built three agents on three models to compare head-to-head — including a benchmark against Amazon Nova — and provisionally selected Haiku, tuned through system instructions. VeUP coached on prompt-trimming and token minimization, temperature settings, and trace-based per-step latency analysis. Two agents are in active development: an AI product-question chatbot with session continuity, and an onboarding agent that operates the admin console rather than emitting static templates.
Production outcomes
| KPI | Result |
|---|---|
| Production outcomes | the customer runs Claude-on-Bedrock agents with guardrail-enforced prompt-injection protection and a benchmarked model-selection rationale (Haiku for latency-sensitive paths, Sonnet for deeper reasoning), arrived at by comparing three live agents on three models. VeUP delivered a Well-Architected Review of the production workload (March 2026) and runs an ongoing managed-billing and technical/alliance advisory relationship governing the agentic roadmap, model selection, and guardrail posture. Quantified per-agent latency and cost are early-stage and available to AWS on request rather than published. |
| Engagement window | 2024-03-31 (Managed Billing relationship); agentic build advisory through 2025 → Well-Architected Review delivered March 2026; agentic build in active development; ongoing managed-billing + advisory |
| Cost / TCO posture | Managed-billing / FinOps relationship governs cost; VeUP coached on token minimization (prompt-trimming) and model cost-routing (Haiku for latency-sensitive steps, Sonnet reserved for deep reasoning) as the primary agent-cost levers. Per-agent cost figures early-stage; not published. |
| Lessons & continuation | For a credential-security ISV, Bedrock Guardrails prompt-injection hardening is the load-bearing control and must be designed before the agent ships; a three-agent / three-model head-to-head is the defensible way to choose between Claude tiers and Amazon Nova; reserve the deep-reasoning model for the steps that need it and route latency-sensitive steps to the cheaper/faster model. |
Amazon Bedrock · Bedrock Agents (session memory + orchestration) · Bedrock Knowledge Bases (RAG) · Bedrock Guardrails (prompt-injection protection) · Anthropic Claude (3.5 Sonnet, Haiku)
Architecture
A net-new agentic AI build on the existing MyCena Desk Center console: Amazon Bedrock Agents orchestrate session memory, Bedrock Knowledge Bases retrieval-augmented generation, and Bedrock Guardrails prompt-injection protection over a benchmarked Claude 3.5 Sonnet / Claude Haiku / Amazon Nova selection — annotated against all six WAFR pillars.
